An Improved Multi-Output Gaussian Process RNN with Real-Time Validation for Early Sepsis Detection

1. Introduction

• Sepsis is difficult to identify early because its symptoms overlap with other conditions, while delayed treatment is associated with greater mortality risk.The introduction reports a 7.6% increase in septic-shock mortality risk for every hour treatment is delayed after hypotension begins.Source: §1 · p. 1
• Existing EHR warning scores are broad, use few variables and only recent values, and may trigger many cancelled alerts.The NEWS alert previously implemented at the authors' health system had 63.4% of alerts cancelled by receiving nurses.Source: §1 · p. 2
• In one ICU case, the model detected sepsis 17 hours before antibiotics and 36 hours before the sepsis definition was met.The patient was deteriorating after invasive cardiac surgery.Source: Figure 1 · p. 2
• Retrospective sepsis labels are noisy because onset is not directly measurable, and clinical measurements are irregular and informatively missing.Sepsis is inferred from abnormal labs and vitals, antibiotics, and blood cultures.Source: §1 · p. 3
• The proposed model uses an MGP to denoise and impute physiological time series, with medication-dependent means, then feeds latent values to an RNN classifier.The RNN also uses informative missingness patterns; training and validation use 18 months of heterogeneous inpatient encounters and include a new real-time scheme.Source: §1 · p. 3

2. Related Works

• NEWS and MEWS assess general clinical deterioration, while SIRS, SOFA, and qSOFA are sepsis-oriented screening scores.The passage notes that SIRS has been phased out in recent years in favor of sepsis-designed scores.Source: §2 · p. 3
• Prior machine-learning work includes Cox regression for sepsis prediction and time-series models for broader deterioration or ICU admission.Gaussian-process studies modeled multivariate physiological series primarily for forecasting, rather than event prediction.Source: §2 · p. 3
• Other work applied RNNs to clinical classification and informative missingness, while this paper builds on end-to-end learning of an MGP and classifier.The prior applications cited include diagnosis-code prediction, heart-failure onset, and missingness patterns in physiological ICU series.Source: §2 · p. 3

3. Proposed Method

• The overall schematic feeds clinical-series latent values into the RNN at each grid time, alongside baseline covariates and medication indicators.The proposed model increases flexibility in both the Gaussian-process component and the RNN classifier.Source: Figure 2 · p. 4, §3 · p. 4
• For a newly admitted patient, the goal is to update the probability of future sepsis using available labs, vitals, medications, and baseline information.The task is framed as multivariate time-series classification.Source: §3 · p. 4
• The method builds on an earlier framework that combines a Multitask Gaussian Process with a Recurrent Neural Network classifier.The paper first introduces the original MGP-RNN before describing its extensions.Source: §3 · p. 4

3.1 The MGP-RNN Classifier

• The MGP-RNN handles irregularly spaced, variable-length multivariate series by imputing and denoising them onto a regular grid for an RNN classifier.The MGP supplies a posterior over latent physiological values at shared grid times, retaining uncertainty about those values.Source: §3.1 · p. 4, §3.1.1 · p. 5

3.1.1 MULTI-OUTPUT GAUSSIAN PROCESSES TO DE-NOISE AND IMPUTE

• The multi-output GP models correlations across variables and time, with a separable kernel factoring those covariances into output and time components.The covariance structure uses an output covariance matrix and a shared time correlation function; the paper later relaxes the separability assumption.Source: §3.1.1 · pp. 4–5

3.1.2 LONG SHORT TERM MEMORY RNNS TO CLASSIFY

• At each grid time, the LSTM receives latent physiological values, baseline covariates, and medication counts to learn time-varying interactions.The classifier uses latent values from the MGP posterior, rather than directly observed latent variables.Source: §3.1.2 · p. 5
• Training minimizes expected classification loss over the MGP posterior, using Monte Carlo samples and reparameterized gradients for the RNN and GP parameters.The implementation uses ADAM and the Lanczos method to accelerate sampling from large multivariate normal distributions.Source: §3.1.2 · p. 5

3.2 Increasing Flexibility of the Multitask Gaussian Process

• The MGP’s two flexibility extensions are medication-dependent means and a relaxed separable kernel.The kernel extension uses a sum of separable covariance functions.Source: §3.2 · p. 6, §3.2.2 · p. 6

3.2.1 INCORPORATING MEDICATION EFFECTS

• Medication effects shape each physiological variable’s prior mean according to previously administered drugs.The medication-response function allows effects on different length scales, and each new drug dose produces a mean-function spike.Source: §3.2.1 · p. 6

3.2.2 SUM OF SEPARABLE KERNEL FUNCTIONS

• A sum of separable covariance functions relaxes the shared temporal-correlation structure imposed across output variables.The authors used Q = 3 in practice.Source: §3.2.2 · p. 6

3.3 Improving the RNN Classifier

• The classifier improvements combine target replication with missingness-pattern inputs.Target replication increases the signal near the series end; missingness indicators are appended to the RNN inputs.Source: §3.3 · p. 6, §3.3.1 · p. 6, §3.3.2 · p. 6

3.3.1 TARGET REPLICATION

• Target replication trains the RNN outputs at multiple time points rather than only at the final step.Labels cover times from 2 hours before to 6 hours after a sepsis event, accommodating imprecise event timing.Source: §3.3.1 · p. 6

3.3.2 UTILIZING MISSINGNESS PATTERNS

• Appending binary indicators of recently sampled labs lets the RNN use missingness patterns alongside MGP values, baseline covariates, and medications.Lab ordering can itself signal a suspected problem.Source: §3.3.2 · p. 6

4.1 Data Description

• The cohort comprised 51,697 heterogeneous inpatient admissions over 18 months, with a 21.4% sepsis rate.Sepsis was identified using persistently abnormal vital signs, a blood culture for suspected infection, and an abnormal organ-failure lab.Source: §4.1 · p. 7

4.2 Experimental Setup

• Models were trained on 80% of encounters, with 10% each reserved for validation and final evaluation; RNNs used a two-layer, 64-unit-per-layer LSTM.Training used L2 regularization, early stopping, and minibatches of 100 encounters.Source: §4.2 · p. 7

4.2.1 CASE CONTROL MATCHING

• Each sepsis encounter was matched to four non-sepsis encounters with similar stay lengths and baseline covariates, aligning controls by the case’s fraction of stay.Training and evaluation data extended through six hours after the case’s sepsis time or the control’s matched prediction time.Source: §4.2.1 · pp. 7–8

4.2.2 METHODS COMPARED

• The compared models ranged from Base MGP-RNN through extensions for target replication, sum-of-separable kernels, medication effects, and missingness indicators.Raw RNN, PLR, RF, NEWS, and MEWS were also included as baselines.Source: §4.2.2 · p. 8

4.3.1 MATCHED LOOKBACK VALIDATION

• At 4 hours before sepsis, the best model had about 1.4 false alarms per true alarm at 80% sensitivity, versus 2.2 for Base MGP-RNN and 3.2 for NEWS.The matched-lookback figure reports AU-ROC, AU-PR, and false alarms per true alarm across prediction lead times.Source: Figure 3 · p. 8, §4.3.1 · p. 9

4.3 Evaluation Schemes and Results

• Evaluation used AU-ROC, AU-PR, and false alarms per true alarm, with matched-lookback predictions evaluated up to 12 hours before sepsis or matched prediction time.The lookback scheme measures how performance changes with prediction lead time.Source: §4.3 · p. 9, §4.3.1 · p. 9

4.3.2 REAL-TIME VALIDATION

• In real-time validation, the proposed model substantially reduced false alarms per true alarm relative to NEWS across sensitivities.This comparison used the best model, Missing indicators, and NEWS.Source: Figure 4 · p. 9, §4.3.2 · p. 10

• The proposed methodology outperforms strong baselines and clinical benchmarks and substantially improves on the related previous model.Source: §5 · p. 10
• Future work includes improving score interpretability, modeling clinical subpopulations, and accounting for uncertainty in classifier parameters.Source: §5 · p. 10
• The authors suggest combining detection with reinforcement learning for treatment strategies and adapting the methods to other adverse events, including cardiac arrest or ICU admission.Source: §5 · p. 10
• A dashboard screenshot with fake data illustrates plans to display model predictions to a sepsis rapid-response team.The authors say the dashboard is being deployed in hospital wards to facilitate early detection.Source: §5 · p. 10, Figure 5 · p. 10
• The authors anticipate that using the model instead of NEWS could reduce alarm fatigue and provider burden and improve patient outcomes.These are projected benefits, not reported evaluation results.Source: §5 · p. 10